How do tree density and body size influence acoustic signals in Amazonian nurse frogs?

ABSTRACT The Acoustic Adaptation Hypothesis (AAH) predicts that acoustic signals emitted at sites with greater vegetation density should have spectral and temporal characteristics that increase signal transmission, but there is a pleiotropism related to body size: large animals produce signals with lower frequency. We used 238 advertisement calls of 34 populations of Amazonian nurse frogs from two Amazonian rainforests with different vegetation density to test if tree density influences the evolution of acoustic parameters. We used PGLS to test for relationships between acoustic traits and phenotypic, environmental and geographic predictors. Spectral and temporal features of calls have an allometric relationship with body size. We found a novel quadratic relationship between note duration and body size. The allometric relationship between dominant frequency and body size and a direct effect of tree density indicates that the evolutionary trajectories of Amazonian nurse frogs follow a general macro-evolutionary pattern as in birds. The temporal features of calls have opposite evolutionary trajectories to those predicted by AAH; frogs from lower tree density environments emit longer notes and have higher note rates than those from denser-tree environments. Subtle differences between Amazonian forest types can drive acoustic diversification of temporal and spectral features of calls at micro-evolutionary scales.


Introduction
The environment influences characteristics of acoustic signals because physical characteristics (e.g.density of shrubs and trunks, air humidity and presence of waterbodies) affect the propagation of sound waves (Morton 1975;Farina 2013).Forest environments are more prone to attenuation and reverberation of acoustic signals by vegetation than open environments (Morton 1975).According to the Acoustic Adaptation Hypothesis (AAH) developed for birds, reverberation and sound absorption in environments with denser vegetation favour calls with lower frequencies, longer duration and lower note rate (Morton 1975;Wiley and Richards 1978;Richards and Wiley 1980).Although some studies have tested AAH in anurans, there is no consensus regarding the environmental influence on the evolution of acoustic-signal components (Ey and Fischer 2009;Erdtmann and Lima 2013;Goutte et al. 2016Goutte et al. , 2018)).Furthermore, most multi-taxa studies have been carried out with species that occur in strongly contrasting environments, such as open and forested areas (e.g.Zimmerman 1983;Goutte et al. 2018).In Amazonia, Salvático (2014) found a direct effect of continuous changes in vegetation structure on the temporal and spectral traits of acoustic signals emitted by a widely distributed nurse frog.However, the influence of tree density across subtly contrasting environments on acoustic traits of Amazonian anurans at multi-species scale remains unexplored.
Acoustic signals are pleiotropic with body size (Martin and Gans 1972;Gerhard and Huber 2002).Call frequency scale allometrically with body size at large macroevolutionary scales over wide ranges of body-size variation, from anurans (Gerhardt et al. 1996;Gingras et al. 2013;Goutte et al. 2018) and birds (Podos, 2001;Gonzalez-Voyer et al. 2013) to mammals (Barclay and Brigham, 1991;Fitch, 1997;Pfefferle and Fischer 2006).However, whether such large-scale allometric principles occur in the same form at lower phylogenetic levels (e.g.within a genus) has been little studied, especially in genera with small body-size variation (<11 mm).Understanding the role of environment and morphology on acoustic-trait evolution at micro-phylogenetic scales could help to link micro-and macroevolution.
Allobates is a diverse genus of small nurse frogs with more than 50 species distributed mostly in Amazonia (Grant et al. 2017).They are territorial, and most species inhabit the leaf litter of forests with different tree densities.Females are polyandric, and a female may mate with all males in her home range (Ursprung et al. 2011;Souza et al. 2017;Rocha et al. 2018).The average body length is smaller than in most other genera of neotropical anurans and varies between 13.8 and 24.0 mm.Those characteristics make Allobates a useful model for understanding the adequacy of AAH predictions in subtly contrasting environments, such as those in Amazonia.We hypothesised that the variation in forest structure influences the acoustic signals of Allobates, with differences in acoustic signals between those occurring in forests with higher tree density and those occurring in more open forests.
Considering the different intrinsic and extrinsic factors that can influence acoustic communication, we investigated whether the acoustic signals of 21 species of Allobates (represented by 34 populations with body lengths ranging from 13.8 to 20.3 mm) in the Amazonia have been shaped by body size and local adaptation, independent of geographic distance.We hypothesised that for these species with little size variation, the temporal and spectral acoustic traits would be not related to species body size, as predicted by AAH, which predicts that environments with denser vegetation favour species with lower frequencies, longer note duration and lower note rate.

Study area
The study was conducted in 10 locations in Brazilian Amazonia (Figure 1), six in Dense Rainforest (FOD) and four in Open Rainforest (FOA) (IBGE 2010).FOA is a forest with lower tree density, which have been characterised as a transition type between Amazonian and extra-Amazonian forests (Veloso et al. 1991).FOD has trees of relatively uniform height that reach up to 50 m, and FOA is characterised by a predominance of smaller plants, such as palms, vines, Strelitziaceae and bamboos (IBGE 2010).

Study organisms
We used advertisement calls of 238 individuals representing 34 populations of 21 species of Allobates (Anura: Aromobatidae) distributed in 10 localities.In each location, 2 to 4 species were recorded (Table 1).Among the 21 recorded species, 13 were restricted to one location and 8 occurred in 2 or more sampled locations.However, 20 of those species occurred exclusively in one forest type, except for A. nidicola, which occurred in both FOD and FOA.The mean body size (snout -vent length -SVL) of the species ranged from 13.8 to 20.3 mm.

Bioacoustic sampling
Calls were recorded by AP Lima during the rainy season between 2002 and 2020, which coincides with the reproductive period of the studied species.The advertisement calls of active males were recorded in the morning and late afternoon, which represent the species vocal-activity peaks.Recordings were made using an AKG 568 EB microphone connected to a Sony WM-D6C recorder.The microphone was positioned approximately one metre from each recorded male.The air temperature was measured with a digital thermometer, and the SVL of recorded males was measured with a digital calliper.
The number of recorded males per population ranged from 4 to 15, totalling 238 recordings.The following parameters were measured: note duration, note rate and dominant frequency.Acoustic parameters were measured using Raven (Charife et al. 2010), following the configuration proposed by Kaefer and Lima (2012): window = Blackman; Discrete Fourier Transform = 2.048 samples and 3 dB filter bandwidth = 82 Hz.Peak frequency was measured by the peak frequency command.
The advertisement calls of the 21 species of Allobates analysed in the present study can be classified into two types (Figure 2): calls composed of a single continuous note (A. caeruleodactylus, A. magnussoni, A. masniger, A. nidicola e A. subfolionidificans) and calls composed by series of two or more notes (A. caldwellae, A. carajas, A. crombiei, A. flaviventris, A. aff. gasconi 1, A. aff. gasconi 2, A. grillicantus, A. grillisimilis, A. nunciatus, A. paleovarzensis, A. tapajos, A. kamilae, A. tinae, A. aff.tinae 1, A. aff.tinae 2 e A. aff.trilineatus 1).For species with calls characterised by the continuous emission of notes, the acoustic parameters were measured at 20 notes per individual, distributed over 1 min of recording.For species with calls composed of a series of notes, the notes were measured in five calls.For both types of calls, note rate was calculated by counting the total number of notes emitted within the 1-min interval.Acoustic parameters were summarised using averages per individual, at each location, giving summarised data for 238 individuals (see Table 1).

Vegetation sampling
Initially, we used Amazon biomass data available from Baccini et al. (2012), extracted for each location using QGIS (version 3.4.132020).In the study sites, Open Rainforest (FOA) had biomass ranging between 197 and 220 mg/Ha, while Dense Rainforest (FOD) had biomass ranging from 274 to 355 mg/Ha.In a study of tree density along the Purus-Madeira interfluve, a transition area between Open and Dense Ombrophilous Rainforest, Schietti et al. (2016) demonstrated that forests with higher biomass tend to have more trees, so they likely have greater absorption of acoustic waves.However, as those data were collected over almost two decades, biomass may have varied over these years.Following the recommendation of Hardt and Benedict (2020), we tested whether the two categories of forests were structurally different.Biomass values of each location were compared between forest types using a t-test, which indicated similarity between categorical and continuous data (t-test: t = −14.1,df = 27.6,p < 0.001).We compared these two forest categories, which are structurally distinct, determined by the IBGE vegetation map (2010), since we did not have access to vegetation data for the same years as the recordings.

Statistical analysis
Acoustic signals emitted by anurans may be under distinct and non-exclusive selective pressures (Ziegler et al. 2011;Farina 2013;Escalona-Sulbarán et al. 2019), such as environmental variation, body size and phylogeny.Investigating the role of environment and body size in the evolution of acoustic signals without considering the effect of phylogeny can bias the results and conclusions about the AAH.However, this type of bias can be minimised or eliminated using comparative phylogenetic methods.
We investigated the determinants of acoustic variables by considering every individual (n = 238) from the 34 populations of the 21 species of Allobates as a sampling unit.We used linear models estimated through Generalised Least Squares (GLS) to test for relationships between each acoustic trait (dominant frequency, note duration and note rate) and phenotypic, environmental and geographic predictors, while accounting for phylogenetic autocorrelation (Paradis and Schliep 2019).Acoustic traits were logtransformed to account for heteroscedasticity.For each acoustic trait, we used the following predictors: body length, habitat (FOA and FOD), temperature and site latitude and longitude to account for potential geographic gradients independent of body size and habitat.
We began by considering a more complex model including a squared body size term to test for non-linear scaling of acoustic traits and a body size × habitat interaction term to test for different acoustic scaling between habitats.If these terms were not supported statistically (P < 0.05), they were excluded, and a simpler model testing for linear scaling and/or habitat-independent scaling was fitted for the final model.Model explanatory power was measured by the squared correlation between observed and predicted values (R 2 ), and phylogenetic relationships were controlled by imposing a phylogenetic structure on model residuals given by Pagel's lambda (a measure of residual variance explained by phylogenetic relatedness, typically varying between 0 and 1) (Paradis and Schliep 2019).To visualise model results, we used partial residuals of acoustic traits to depict the relationship between a given acoustic trait and a given predictor while controlling for the remaining predictors in each model (Breheny and Burchett 2017).Statistical analyses were conducted using R 4.1.1(R Core Team 2019) with the packages 'ape', 'phytools', 'caper', 'nlme' and 'visreg'.
Additional 16S rRNA sequences from other nominal and candidate Allobates species obtained by other studies were downloaded from GenBank, as well as a representative sequence from Anomaloglossus stepheni, used as an outgroup.The database containing all sequences was aligned following the standard configuration of the MAFFT online implementation (Madeira et al. 2019), except for the use of the E-INS-i strategy.The resulting alignment was edited in Geneious 5.3.4.The final alignment consisted of 87 terminals and 556 base pairs.The best explanatory evolutionary model for final alignment nucleotide substitutions was verified using PartitionFinder 2.1.1 (Lanfear et al. 2016) via CIPRES (phylo.org),using Bayesian Inference Criterion and PhyML algorithm (Guindon et al. 2010); GTR+G+I was retrieved as the most suitable evolutionary model.
The phylogenetic relationships of the 21 species included here and other congeners were inferred through Bayesian inference with BEAST 2.5 (Bouckaert et al. 2014), configured as follows: Yule tree prior, strict clock model, single run of 100 million sampled generations every 100 million sampled generation every 10,000 Markov chain Monte Carlo chain (MCMC).The stationary of the posterior distributions of all parameters was verified using Tracer 1. 7 (Rambaut et al. 2018); all parameters had an estimated sample size (ESS) greater than 200.The maximum clade credibility tree (MCC) was reconstructed using TreeAnnotator (Rambaut et al. 2018) after 10% burnin.The MCC containing all terminals was reduced by using the 'drop.tip'command of the package 'ape' (Paradis and Schliep 2019) to contain only the 21 species included in the present study (Figure 3).Finally, polytomic branches were artificially added to the phylogenetic tree in order to represent each of 238 individuals.

Results
The phylogenetic relationship of the 21 species of Allobates inferred through Bayesian inference is similar to recently published phylogenies.One large clade is composed only of dense rainforest species; all other large clades are composed of both open-and denserainforest species (Figure 3).

Effects on spectral parameters
The Generalised Least Squares (GLS) model considering phylogeny as a co-factor indicated a significant association of dominant frequency with forest type (T = −2.30,P < 0.022), body size (T = −4.12,P = 0.0001) and latitude (T = 2.55, P = 0.011), but not with longitude (T = −0.79,P = 0.425) or temperature (T = 0.12, P = 0.899).The squared correlation between the observed and predicted values was high (R 2 = 0.71).Dominant frequency was higher in FOA than in FOD (Figure 4(b)), decreasing with body size increase (Figure 4(a)) and increasing to the north towards the Amazon River lowlands (Figure 4(c)).

Effects on temporal parameters
We tested note rate and note duration in separate models.The GLS model considering phylogeny as a co-factor revealed that note duration is affected by forest type (T = −3.78,P < 0.001), body size (T = 4.54, P < 0.001) and latitude (T = 2.47, P = 0.014), while temperature (T = −1.92,P = 0.055) and longitude (T = 0.67, P = 0.501) showed no arbitrary statistical effects, but the probability of no effect of temperature was small.The squared correlation between the observed and predicted values was low (R 2 = 0.28) compared to the dominant-frequency model.Note duration showed a quadratic relationship with body size, with longer duration in intermediate-sized individuals than in smaller and larger ones (Figure 4(d)).Furthermore, note duration was longer in FOA than in FOD (Figure 4(e)) and increased to the north towards the Amazon River lowlands (Figure 4(f)).
The GLS model revealed that the note rate is statistically significantly affected by the type of forest (T = 4.26, P < 0.001), body size (T = 3.55, P < 0.001) and latitude (T = −3.18,P = 0.002).Temperature (T = −1.20;P = 0.230) and longitude (T = −1.91,P = 0.057) did not reach an arbitrary level of statistical significance, but the probability that there was no effect of longitude on note rate was small.The squared correlation between the observed and predicted values was the lowest among the three models (R 2 = 0.18).The relationship between note rate and body size was positive in both forest types and higher values occurred in FOD (Figure 4(g)).Note rate decreased to the north towards the Amazon River Lowlands (Figure 4(h)) and tended to decrease towards the east (Figure 4(i)).A summary of the tree linear models estimated through Generalised Least Squares (GLS) accounting for phylogenetic autocorrelation used to reduce bias of phylogeny is shown in Supplementary Material.

Discussion
Our results show a similar spectral-trait pattern to that predicted by the Acoustic Adaptation Hypothesis (AAH) as defined by Morton (1975), but with an opposite temporal pattern.According to the AAH, reverberation and sound absorption in environments with denser vegetation should favour calls with lower frequencies, longer note duration and lower note rate (Morton 1975).However, our results demonstrated that frogs from forested environments with lower tree density emitted longer notes and had higher note rates than those from more densely treed environments, as opposed to the pattern predicted by AAH.The dominant frequencies followed the AAH predictions, with frogs in forests with higher tree density having lower dominant frequency.Most studies that tested AAH in anurans used macro-evolutionary scales (between genera) and did not find the effects of vegetation on temporal acoustic signals in anurans from Asia (Goutte et al. 2016(Goutte et al. , 2018)), Amazonia (Erdtmann and Amézquita 2009) or Atlantic Forest (Bezerra et al. 2021).This leads us to suggest that the evolutionary trajectories of anurans do not follow the same general macro-evolutionary pattern as in birds (Morton 1975).Relationships between vocalisation parameters and the predictors variables.We show only the significant relationship.Regressions from PGLS results are represented by solid lines.These relationships are without the effect of the other variables included in the models.Note that the body size increased as the dominant frequency decreased (a), that was higher in FOA than in FOD (b) and increased to the north (c).Log Note duration showed a quadratic relationship with body size (d), was longer in FOA than in FOD (e) and increased to the north (f).The relationship between note rate and body size was positive in both forest types, higher values occurred in FOD (g), decreased to the north (h) and tended to decrease towards the east (i).
The dominant frequency of advertisement calls is an important property associated with intrinsic characteristics of individuals and species since it depends on the morphology of the sound-producing apparatus and on the neurological system (Wallschlager 1980;Ryan 1986).The correlation between tree density and dominant frequency in this study may be explained by the conserved anatomy of the vocal tract of closely related species (Cocroft and Ryan 1995;Panhuis et al. 2001).Structures that vibrate in the vocal apparatus are known to increase with body size, leading to a negative relationship between size and frequency.However, the variance explained by the pleiotropic connection between call frequency and body size (Martin and Gans 1972;Gerhard and Huber 2002) in multi-taxa studies was smaller (15% to 66%) (Boeckle et al. 2009;Erdtmann and Amézquita 2009;Escalona-Sulbarán et al. 2019;Sugai et al. 2021) than in our study, where body size explained 71% of dominant-frequency variation, despite the low variation in size .High tree density also causes sound attenuation, and this has been reported as an important constraint on the evolution of spectral acoustic signals in tropical birds (Weir et al. 2012).The correlations between dominant frequency and geographic distance (latitude) in our study were probably associated with neutral evolutionary processes, such as genetic drift, as suggested by Pröhl et al. (2007) andAmézquita et al. (2009).
In Asian anurans living near noisy (torrent) environments, background sound was the main driver of the change in spectral traits of calls but had no effect on temporal traits (Goutte et al. 2018).In Amazonia, after controlling for phylogeny, frog species in noisy environments tend to be smaller (Vargas-Salina and Amézquita 2013).Apparently, on a macro-evolutionary scale, environments with physical noise drive the spectral characteristics of sound in Asian (Goutte et al. 2018;Zhao et al. 2021) but not in the neotropical forest (Vargas-Salina and Amézquita 2013).Noise produced by other species was not considered to be the main driver of multi-taxa acoustic adaptation (Amézquita et al. 2011).Overall, our study of neotropical frogs within a single genus suggests little variation in dominant frequency, but strong variation in temporal variables.In contrast to other studies, we found a strong effect of tree density on these variables, independent of body size, phylogenetic signal, distance and temperature.Note duration was shorter, and the note rate was higher in sites with more trees.We suggest that adaptation at the micro-evolutionary scale affects call temporal variables more than dominant frequencies because the small variation in body size results in a narrow frequency spectral window, but the call temporal structure provides a huge number of possible combinations, with single or double notes emitted in series (from 2 to 63 notes) or continuously.This enables a large variety of calls and allows the co-existence of many species despite the limited spectral window.
We believe that the relationship between body-size and temporal traits of calls are determined at micro-evolutionary scales, and this may explain why such relationships have not yet been observed at macro-evolutionary scales for neotropical (García et al. 2014;Mason and Burns 2015;García and Tubaro 2018;Bezerra et al. 2021) or Asian frogs (Goutte et al. 2016(Goutte et al. , 2018)).Temporal traits of calls are more easily modifiable at the species or individual levels (Cocroft and Ryan 1995), but the quadratic pattern found in note duration indicates physiological or neurological restrictions on continuous increase in a single temporal characteristic of calls at micro-evolutionary scales.
Note rate increased with body size, which was an unprecedented relationship in frogs, as opposed to what is known for birds (Neotropical cardinalids and Thraupids), where smaller animals have higher note rate (García et al. 2014;Mason and Burns 2015).The reason behind this pattern is not clear, but it may be related to morphological characteristics or to the sound-emitting apparatus, which differs between birds and frogs.In anurans, the larynx muscles are simpler than in birds (Wells and Schwartz 2007).Bird species for which such a relationship has been registered belong to the group known as 'songbirds', which have good control of the syrinx muscles and can generate complex sounds (Suthers et al. 1999;Riede and Goller 2014).
In this study, we demonstrated that the characteristics of advertisement calls of Allobates species were determined by body size, environmental adaptation and other unknown evolutionary forces.We compared phylogenetically closely related tropicalforest species and show that, in addition to phylogenetic proximity, body size and subtle differences in tree density also lead to diversification in the acoustic parameters at microevolutionary scales (within genus).We suggest that more studies at micro-evolutionary scales should be carried out on tropical-forest anurans to better understand the factors that influence acoustic parameters, as tropical forests retain the world's greatest diversity of anurans.

Figure 1 .
Figure 1.Map with study sites and their respective forest classification, according to IBGE (2010).

Figure 2 .
Figure 2. Types of call structure in Allobates.(A) Calls composed of a single note, continuously emitted (A.caeruleodactylus, Borba).(B) Calls composed of groups of two or more notes (A.nunciatus, Trairão).

Figure 3 .
Figure 3. Maximum clade credibility 16S phylogenetic tree summarising the relationship of 21 species of Allobates used as input in the phylogenetic generalised least squares analyses.Coloured squares denote the forest types.Scale bar represents nucleotide substitutions per site.Sampling localities: AUT, Autazes; BOC, Boca do Acre; BOR, Borba; BRN, Brasil Novo; MAD, east bank of the upper Madeira River; MAE, west bank of the upper Madeira River; PNA, Amazonia National Park; RBR, Rio Branco; TRA, Trairão; TRE, Treviso.

Figure 4 .
Figure 4. Relationships between vocalisation parameters and the predictors variables.We show only the significant relationship.Regressions from PGLS results are represented by solid lines.These relationships are without the effect of the other variables included in the models.Note that the body size increased as the dominant frequency decreased (a), that was higher in FOA than in FOD (b) and increased to the north (c).Log Note duration showed a quadratic relationship with body size (d), was longer in FOA than in FOD (e) and increased to the north (f).The relationship between note rate and body size was positive in both forest types, higher values occurred in FOD (g), decreased to the north (h) and tended to decrease towards the east (i).

Table 1 .
Allobates species sampled and the number of males recorded by location and forest type.Geographical coordinates are displayed in decimal degrees.Abbreviations: FOA = Open Rainforest; FOD = Dense Rainforest.